Giving an agent continuity
Letta builds software for stateful agents: applications that retain memory and other state as interactions continue. Its September 2024 introduction described a model-agnostic layer above inference, so an application's stored information could remain useful when developers changed model providers. The company puts memory management and agent execution at the center of that design. [1]
This treats an agent as a continuing service, with tool use and external data as well as chat messages. Managing the context supplied to a model becomes part of the application, rather than leaving all continuity to the model provider. Letta's announcement explains that architectural aim; it does not demonstrate that the stored memories are always correct or relevant. [1]
MemGPT is the research, Letta is the company
Letta publicly introduced itself on September 23, 2024, continuing work from the MemGPT project at UC Berkeley. A companion post reserved MemGPT for the original agent design pattern with self-editing memory tools and used Letta for the agent framework. The company said it would continue maintaining the open-source software. [1][2]
At the introduction, an Agent Development Environment was still being developed and Letta Cloud was entering a pilot. Those plans are distinct from the already available open-source project the company was taking forward. [2][1]
Sources
- Announcing Letta
Letta · Sep 23, 2024
- MemGPT Is Now Part of Letta
Letta · Sep 23, 2024